Author: Sidney Ocanagil-Tunstall
Berlin, European Society of Medical Oncology 2025

The change didn’t go unnoticed. Rather than continuing where Dr Miere Crispin Ortuzar left off, Dr Inti Zlobec steered the discussion from the Current Use of AI in Pathology toward a tougher question: why, with all the technology in place, is full-scale AI implementation still out of reach?

Dr Crispin Ortuzar had earlier provided an introductory overview of AI in digital pathology image analysis during her session. But Dr Zlobec’s new title read: Why is pathology so slow at implementing AI. Clearly, this was the sticking point Dr Zlobec wanted to unpack.

The transport is ready, but we’ve got no roads to drive on

It’s a bit like having a fleet of high-performance engines and nowhere to race them. The technology exists, the algorithms are ready, yet we do not see worldwide adoption, or even individual nations, ahead on the digital healthcare curve, racing to a digital finish. In Switzerland, “we estimate only 5% of pathology labs are paperless, working in a fully digital way.”

Dr Zlobec began by asking: Are pathologists ready to use AI in practice? Her “reality check” was blunt. Although pathologists are ready for AI uptake, she argued that “only a fully digitised lab can fully benefit from the potential of AI and do so without suffering from logistical challenges.”

Partial digitisation can create chaos. “The lab must know that some slides need to go for scanning,” she asked, “while another set of slides does not” Without a complete digital workflow, this hybrid approach can become a logistical nightmare. The only viable route to take complete advantage of AI solutions, she explained, is full digitisation: “Trust me, we’ve tried.”

The bottleneck beneath the microscope

So, what’s really holding back implementation? According to Dr Zlobec, it’s in-part not the algorithms, but the systems that connect them.

“This [going digital with the intention of adopting AI] is a major IT project… with the biggest problem of all being the basic interface between the Laboratory Information System (LIS) and Image Management System (IMS)… to be honest, I’ve never seen or heard of this working seamlessly in any lab in Europe or across the world.”

Even if a lab wants to make the leap, many simply can’t. “You might have to rebuild your lab.” Digital pathology labs, she reminded the audience, are built different and you can’t expect pathologists to get on board with a complete rebuild right away.

“Pathologists have been doing this [traditional pathology] for 200 years… and you’re asking them to go from this to something like the PathoJet.”

The PathoJet is a gaming chair for digital pathologists. “We have a few in our lab.” But adopting digital pathology is about more than hardware, it’s a shift in mentality, in workflow, and in identity. Historically, pathologists have tended to adopt major changes only when circumstances left little alternative. The COVID pandemic, for instance, showed just how quickly transformation could happen when it had to.

But even beyond infrastructure, another constraint looms. What about cost?

A puppet master at large?

Real time conference surveys conducted in rooms packed with pathologists show financial barriers continue to dominate audience polls. “Each scanner costs 250,000 Swiss francs and you might need six in your lab, if your lab is a moderately sized institute like ours.” The initial investment, she added, is around five million euros.

Only once a lab is fully digitised can AI even enter the conversation. But then comes the next dilemma. Which AI?

For those at home in this domain, the exponential growth of image analysis companies has made that question harder, not easier. “In German, we have the phrase, ‘die Qual der Wahl’, the problem of having too many appealing choices.” Many of these AI tools are what she calls “repetitive ,” models boasting the same offering, with numerous Ki-67 scoring algorithms for breast cancer and many Gleason grading algorithms for prostate cancer.

Beyond redundancy, the available algorithm portfolios are small, and most are still research use only (RUO). “Most European countries, with perhaps rare exceptions, don’t offer reimbursement for these algorithms at this point in time.” Without reimbursement, labs have little financial incentive to adopt advanced digital tools.

In light of Paige’s recent acquisition by Tempus, Dr Zlobec also questioned how labs are expected to commit to a single AI company. It’s a “dynamic industry” and understandably labs want to be able to invest in an algorithm developer that has their needs at heart and will continue to do so in the future.

She also challenged the AI industry directly.

“Right now, AI isn’t made all that much for pathologists’ needs… there’s a misalignment between what the pathologists want and if I dare say it, where the money comes from.”

If AI were truly built for pathologists, she argued, “you’d be finding algorithms for quantification of eosinophils, quantification of lymphocytes, and anything to do with counting.” These tools could transform day-to-day diagnostic work and streamline repetitive tasks allowing pathologists to focus on the complex ones.

Then there’s the issue of interoperability. She cited the example of up-and-coming computational companion diagnostic tests for patient prediction, and challenges arising when the test can only be used when the stainer, scanner, and software all come from the same vendor family.

The Closing Statement

Looking ahead, Dr Zlobec turned to cytopathology, a field already showing “amazing advancements” in its pursuit of AI implementation. Perhaps digital pathology could learn a trick or two from this speciality?

Bringing us full circle, she noted, “It’s not always about resistance from pathologists.” “Pathologists have [consistently] shown their adaptability” when confronted by new technologies. But, “funding and software integrations continue to slow the pace of progress. Pathologists want validation, quality assurance, patient safety, and accountability”.

Finally, Dr Zlobec left the audience with a question to ponder on, “If pathology labs cannot supply these computational diagnostic tests, what’s going to happen?” Now that we are experienced, we know the terms, while we witness cancer rates rise, and our capacity decline. Is it now not more fundamental than ever, that we look to embrace and implement AI in the lab.